Re: [PATCH] anon_inodes.c cleanups.

[Rusty Russell <rusty@rustcorp.com.au>]

On Saturday 12 April 2008 11:15:26 Davide Libenzi wrote:
> On Fri, 11 Apr 2008, Rusty Russell wrote:
> > Arnd pointed me at anon_inode_getfd(), and the code annoyed me enough
> > to send this patch.
> >
> > Mainly because the init routine carefully checks for errors, then panics
> > (because we shouldn't run out of memory at boot).  Unfortunately, it's
> > actually worse than simply oopsing, where we'd know what had failed.
> >
> > 1) anon_inode_inode can be read_mostly, same as anon_inode_mnt.
>
> Sure.
>
> > 3) anon_inode_mkinode has one caller, so it's a little confusing.
>
> Hmm? The function groups the code necessary to create the anonfds inode.
> If every function that has one call site would be inlined, we'd have
> monster long functions. Functions also have the purpose to group some code
> that does some task, into a single unit (and the function name hopefully
> gives an hint about what's doing). The compiler (not that in this case
> really matter, since it's not even a slow-path, is a once-run path) may
> take care of inlining, if sees that appropriate.

If you'd called it, say, "setup_anon_inode()", it would be fine.  It seems 
overly generic unless you planned on calling it elsewhere.

> > 2) The IS_ERR(anon_inode_inode) check is unneeded, since we panic on
> >    boot if that were true.
> > 4) Don't clean up before panic.
> > 5) Panic gives less information than an oops would, plus is untested.
>
> I remember we changed the failure-path of anonfds a couple of times along
> the way, but I can't find email traces about why we did it.
> So, I prefer error-checked code instead of oopses, and given that the
> anonfds subsystem is not a required one for most of the components of the
> kernel/userspace, I'd rather prefer to drop the panic().

We've seen this debate before, and I'm firmly on the "don't turn oopses into  
errors on boot paths" side.  I know others disagree.

Given that it should never happen, I'd argue the highest priority minimal 
amount of code, and second is ease of debugging if it ever did happen to 
someone.  Oopsing has those features.

> Anyway, I'll let this handle with Al (cc-ed now). The ananofds interface
> has been changed to remove the inode** and file** parameters (noone but
> KVM was using them), and Al already has those changes in his vfs tree
> (plus fixes for KVM, I think).

OK, fine.

Thanks,
Rusty.